SpaceXAI releases flagship Grok 4.6 model with advanced reasoning capabilities
SiliconANGLE Maria Deutscher ● Covered by 3 sources
SpaceXAI just shipped Grok 4.6, its new flagship AI model. It says the longer training helped it match GPT-5.6 Sol and edge closer to Claude Fable 5.
Based on reporting by SiliconANGLE, Maria Deutscher — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
SpaceXAI has released Grok 4.6, its new flagship large language model, and says the update pushes deeper into reasoning, coding and science work. The company claims it can beat Anthropic’s Claude Fable 5 in some areas, while landing on par with OpenAI’s GPT-5.6 Sol on one of its main scores.
The launch comes only a month after SpaceXAI’s last flagship model. The company says the big difference this time is simple: it spent longer training the model. That extra run used an AI-generated dataset meant to sharpen reasoning, plus “high-quality engineering data” pulled into the process.
After that first pass, engineers ran two more steps. One was supervised fine-tuning, the familiar stage where a model is taught from sample prompts and prepared answers so it gives cleaner, more usable replies. The other was reinforcement learning. SpaceXAI also used Grok 4.5 to help optimize Grok 4.6’s supervised fine-tuning phase, with a focus on science and programming tasks.
For its own evaluation, SpaceXAI used the Artificial Analysis Intelligence Index, which combines nine popular AI benchmarks across areas like science, coding and financial services. Grok 4.6 scored 61, which the company says puts it one point behind Claude Fable 5. SpaceXAI also said Grok 4.5 beat Claude Fable 5 in three of nine other benchmarks, including AA-Briefcase, a test for knowledge-work projects that can take a human weeks to finish.
The new model is already being sold: the standard version costs $2 per million input tokens and $6 per million output tokens, while a faster edition costs twice as much. It’s available through Cursor, the vibe coding platform SpaceXAI bought for $60 billion in June, and through Grok Build, the company’s own programming tool. That is a very Musk way to ship an AI product: sell the model, wrap it in tools, and let the benchmarks do the talking.
My take — AI-written commentary, not fact-checked reporting
This is the usual AI arms race move: more training, bigger claims, fresher benchmarks, repeat. The interesting part isn’t the brag sheet, it’s how quickly these model launches are turning into product packaging for developers and coding tools. Everyone loves talking about “reasoning” until the bill arrives in tokens.
Read more about this at: SiliconANGLE
